Research map / Label-efficient and robust learning / Knowledge distillation
Foundry: Distilling 3D Foundation Models for the Edge
Guillaume Letellier, Siddharth Srivastava, Frederic Jurie, Gaurav Sharma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill one generalist student instead of task-specific specialists | Foundry · ShapeNet55 | accuracy (after fine-tuning): 75.09* -> 89.95 | Improved |
| replace learnable DSO with static K-Means (ablation) | Foundry · ShapeNet55 | accuracy: 89.68 -> 76.08 | Hurt |
| use FPS pre-sampling instead of learned compression (ablation) | Foundry · ShapeNet55 | accuracy: 89.68 -> 87.56 | Hurt |
| unfreeze student encoder during distillation (ablation) | Foundry · Distillation loss | distillation loss: 0.1006 -> 0.0736 | Improved |
| soft attention assignment instead of hard argmax (ablation) | Foundry · Distillation loss | distillation loss: 0.0736 -> 0.0787 | Hurt |
| add QKV biases in DSO (ablation) | Foundry · Distillation loss | distillation loss: 0.0736 -> 0.0763 | Hurt |
| add output layer norm at CAU (ablation) | Foundry · Distillation loss | distillation loss: 0.0736 -> 0.0748 | Hurt |
| increase number of SuperTokens beyond 4 (ablation) | Foundry · ModelNet40 and others | accuracy: 91.25 -> 91.41 | No clear effect |
| combine FMD distillation with ToMe | ToMe (s=16) · Avg over benchmarks | average score: 73.80 -> 86.52 | Improved |
| combine FMD distillation with PatchMerger | PatchMerger (s=16) · Avg over benchmarks | average score: 87.59 vs teacher 87.72 | Improved |
| compress to a single SuperToken (s=1 versus s=16) (ablation) | Foundry · ModelNet40 5-way 10-shot | accuracy: 92.5 -> 91.8 | No clear effect |
| increase gate regularization lambda_gate (ablation) | Foundry-Gate · ModelNet40 | Spearman correlation with performance: -0.45 (frozen), -0.13 (unfrozen) | Hurt |
Closest papers by meaning
- SAM-Guided Masked Token Prediction for 3D Scene Understanding (NeurIPS 2024 · similarity 0.87)
- CustomKD: Customizing Large Vision Foundation for Edge Model Improvement via Knowledge Distillation (CVPR 2025 · similarity 0.86)
- Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning (NeurIPS 2025 · similarity 0.86)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.85)
- Point Cloud Dataset Distillation (ICML 2025 · similarity 0.85)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
